Somewhere in France this autumn, supermarket managers signed scrap dispositions on a scale nobody trained them for – and no quality department was in the room. Hackers hit Intermarché. Refrigeration went down. The cold chain disposed of itself: pallets of chilled product into the skip because nobody could prove what they had been through. I have signed scrap decisions on plant floors for two decades, and the ones that keep me awake are never the big ones. They are the ones where the data telling you what to scrap cannot be trusted. That same autumn, the NSA told operators of OT systems that perimeter trust is finished. Here is the uncomfortable observation: your PFMEA lists a power cut and a low air-pressure alarm. It does not list a falsified SCADA value.

The network is a process input, not an IT problem

Walk any modern line and count what arrives over a network: torque results from DC tools, oven and furnace curves, CMM reports, label prints, traceability scans, vision verdicts. The quality system signs all of it. Every byte travels through infrastructure an adversary can touch. At Intermarché the refrigeration telemetry was effectively the quality record for the cold chain – once someone sits in that path, you cannot disposition product, because the evidence itself is under suspicion.

The standards gesture at this. IATF 16949 and AS9100 both make you assess the context of the organisation and the infrastructure your processes depend on; almost nobody reads those clauses as cyber clauses. Validation has the same blind spot. A validated process state assumes the data describing that state is honest, and that assumption is now negotiable – while the name on the release certificate is not the IT department's. It is yours.

This is not getting simpler. Half the trade press celebrates AI agents that adjust setpoints and release work orders on their own. I build and deploy multi-agent AI systems myself; I would not let one near a release decision without the controls I demand from a human operator. Every agent is another writer into your process record.

Calibration certifies the instrument. Nothing certifies the network between the instrument and the record.

Zero trust is a control plan for packets

Strip the NSA's zero-trust guidance of its jargon and it reads like a control plan for packets. Verify every interaction before granting it – a detection control, inspection at every handover. Segmentation is containment: keep one compromised cell from taking the plant. Least privilege is poka-yoke, making the wrong action impossible rather than discouraged. Continuous monitoring is SPC applied to access events.

Quality engineers already think adversarially; we just decline to say so. A PFMEA is a threat model with different paperwork – severity, occurrence, detection instead of impact, likelihood, exposure. I am a certified ethical hacker with thirty-plus security certifications, and the clickjacking flaw that put me in T-Mobile's public bug-bounty Hall of Fame was not found with exotic tooling. It was found by reading a page the way its builders never did: what happens if this element overlays that one? That is a PFMEA question asked about code instead of a stamping die. Nobody asks it about OT, because it appears in no one's job description.

Own the failure mode before it owns you

The RACI matrix has a hole. IT is accountable for the perimeter, quality for the process, and the failure mode between them – adversarial manipulation of process data – belongs to nobody. Adversaries live in exactly that no-man's-land, because nobody patrols it.

Close it with tools you already own. Add data-integrity compromise to the PFMEA wherever a number triggers a release decision; the severity there is brutal – not the part breaking, but losing the ability to prove any part good. Write the reaction plan for the measurement lied: define the suspect window from the last verified checkpoint to detection, quarantine everything inside it, re-verify physically before you scrap.

Then rehearse an OT intrusion the way you rehearse a customer escape. At Witte Automotive our failure-cost reduction came from treating defects as containment-first events – QRQC in the first hour, A3 and 8D on root cause once the bleeding stopped. An intrusion deserves the same sequence: contain the data before you chase the intruder. And put a quality engineer with shipment-stop authority inside the cyber incident team, not on its distribution list.

Run the arithmetic on the cheap side of the argument. The invoice for an intrusion is not the ransom; it is suspect stock. Eight hours of falsified oven logs on an automotive line can quarantine a full shift of production – call it €200,000 of work in progress that ships only after re-verification. That cost lands in the quality budget whether or not anyone planned for it.

The audit question that should retire the firewall checklist

Auditors ask whether you have a firewall. The question that matters is harder: show me where falsified process data appears in your control plan, and walk me through your first hour. Most plants cannot answer, which is at least honest – the answer does not exist yet.

Key takeaways

  • Treat data integrity as a process input: wherever a measurement triggers release, add "adversarial falsification of process data" to the PFMEA with a written reaction plan, not just a severity score.
  • Run an OT intrusion like a customer escape – QRQC containment inside the first hour, a defined suspect window, physical re-verification of quarantined stock, then an 8D on root cause.
  • Translate zero trust into quality vocabulary when you talk to IT: verification is a detection control, segmentation is containment, least privilege is poka-yoke.
  • Give the cyber incident team a named quality engineer with authority to stop shipments – accountability for an unowned failure mode has to sit somewhere.

Hackers and quality engineers practise the same craft: understanding how complex systems fail, then building them so they don't. Only one of the two currently appears on your organisation chart. The network is hostile – that much is settled. The question for the next decade is whether your control plan stays innocent about it, or learns to defend the truth of every number it signs.